Conducts a multi-axis code review of a GitHub pull request and posts it back to the PR as inline review comments plus a summary assessment. Use when asked to review a PR, review a GitHub pull request, leave review comments on a PR, or when a PR number/URL is given for review. For reviewing local uncommitted changes instead, use code-review-and-quality.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Conducts a multi-axis code review of a GitHub pull request and posts it back to the PR as inline review comments plus a summary assessment. Use when asked to review a PR, review a GitHub pull request, leave review comments on a PR, or when a PR number/URL is given for review. For reviewing local uncommitted changes instead, use code-review-and-quality.
PR Code Review and Quality
Overview
Multi-dimensional review of a GitHub pull request across five axes — correctness, readability, architecture, security, and performance — delivered to the PR itself as a single GitHub review: findings anchored inline to file and line, plus an overall assessment in the review body.
The approval standard: Approve a change when it definitely improves overall code health, even if it isn't perfect. Perfect code doesn't exist — the goal is continuous improvement. Don't block a change because it isn't exactly how you would have written it. If it improves the codebase and follows the project's conventions, approve it.
Output rule: The review goes on the PR, not in the terminal. Do not dump the full review as terminal prose. Post it, then report the review URL and a one-paragraph summary of what was posted.
When to Use
The user asks for a review of a GitHub PR (by number, URL, or owner/repo#N)
A PR is ready for merge and needs a quality gate
Another agent or model authored a PR you need to evaluate
Re-reviewing a PR after the author pushed changes
Step 0: Resolve the PR — Required
The PR must be supplied. Accept any of:
123 (bare number — resolves against the repo of the current directory)
owner/repo#123
https://github.com/owner/repo/pull/123
If no PR was provided, stop and ask which PR to review. Do not infer it from the current branch, do not guess from open PRs, do not pick the most recent one. Asking is the only correct move.
Resolve to OWNER, REPO, NUMBER and keep them for every subsequent call.
# Bare number: derive owner/repo from the current directory's remote
gh repo view --json nameWithOwner -q .nameWithOwner
Confirm the PR exists and is readable before reviewing:
gh pr view NUMBER --repo OWNER/REPO --json number,title,state,isDraft,author,baseRefName,headRefName,additions,deletions,changedFiles
Step 1: Fetch the Change
Gather everything before forming an opinion:
# Intent: title, description, linked issues, review history
gh pr view NUMBER --repo OWNER/REPO --json title,body,url,commits,labels
gh pr view NUMBER --repo OWNER/REPO --comments
# The diff — this is what you can anchor inline comments to
gh pr diff NUMBER --repo OWNER/REPO --patch > /tmp/pr-NUMBER.patch
# Per-file stats and patches (useful for large PRs)
gh api repos/OWNER/REPO/pulls/NUMBER/files --paginate
# The head SHA — pin the review to it
gh api repos/OWNER/REPO/pulls/NUMBER --jq .head.sha
# CI state — feeds the verification axis
gh pr checks NUMBER --repo OWNER/REPO
Read beyond the diff. A diff hunk rarely contains enough context to judge correctness. For any file with non-trivial changes, read the full file — from the local checkout if the branch is available, otherwise:
gh api repos/OWNER/REPO/contents/PATH?ref=HEAD_SHA --jq .content | base64 -d
Also read the callers of changed functions, the tests covering them, and neighboring code that establishes project convention. Reviewing a hunk in isolation produces shallow findings.
Large PRs: if the diff exceeds what you can hold usefully, review file group by file group rather than skimming everything. Note in the summary if coverage was partial — never imply full coverage you didn't achieve.
The Five-Axis Review
Every review evaluates code across these dimensions:
1. Correctness
Does the code do what it claims to do?
Does it match the spec, linked issue, or PR description?
Are edge cases handled (null, empty, boundary values)?
Are error paths handled (not just the happy path)?
Does it pass all tests? Are the tests actually testing the right things?
Are there off-by-one errors, race conditions, or state inconsistencies?
2. Readability & Simplicity
Can another engineer (or agent) understand this code without the author explaining it?
Are names descriptive and consistent with project conventions? (No temp, data, result without context)
Is the control flow straightforward (avoid nested ternaries, deep callbacks)?
Is the code organized logically (related code grouped, clear module boundaries)?
Are there any "clever" tricks that should be simplified?
Could this be done in fewer lines? (1000 lines where 100 suffice is a failure)
Are abstractions earning their complexity? (Don't generalize until the third use case)
Would comments help clarify non-obvious intent? (But don't comment obvious code.)
Are there dead code artifacts: no-op variables (_unused), backwards-compat shims, or // removed comments?
3. Architecture
Does the change fit the system's design?
Does it follow existing patterns or introduce a new one? If new, is it justified?
Does it maintain clean module boundaries?
Is there code duplication that should be shared?
Are dependencies flowing in the right direction (no circular dependencies)?
Is the abstraction level appropriate (not over-engineered, not too coupled)?
4. Security
For detailed security guidance, see security-and-hardening. Does the change introduce vulnerabilities?
Is user input validated and sanitized?
Are secrets kept out of code, logs, and version control?
Is authentication/authorization checked where needed?
Are SQL queries parameterized (no string concatenation)?
Are outputs encoded to prevent XSS?
Are dependencies from trusted sources with no known vulnerabilities?
Is data from external sources (APIs, logs, user content, config files) treated as untrusted?
Are external data flows validated at system boundaries before use in logic or rendering?
5. Performance
For detailed profiling and optimization, see performance-optimization. Does the change introduce performance problems?
Any N+1 query patterns?
Any unbounded loops or unconstrained data fetching?
Any synchronous operations that should be async?
Any unnecessary re-renders in UI components?
Any missing pagination on list endpoints?
Any large objects created in hot paths?
Change Sizing
Small, focused changes are easier to review, faster to merge, and safer to deploy. Target these sizes:
~100 lines changed → Good. Reviewable in one sitting.
~300 lines changed → Acceptable if it's a single logical change.
~1000 lines changed → Too large. Split it.
What counts as "one change": A single self-contained modification that addresses one thing, includes related tests, and keeps the system functional after submission. One part of a feature — not the whole feature.
Splitting strategies when a PR is too large:
Strategy
How
When
Stack
Submit a small PR, start the next one based on it
Sequential dependencies
By file group
Separate PRs for groups needing different reviewers
Cross-cutting concerns
Horizontal
Create shared code/stubs first, then consumers
Layered architecture
Vertical
Break into smaller full-stack slices of the feature
Feature work
When large PRs are acceptable: Complete file deletions and automated refactoring where the reviewer only needs to verify intent, not every line.
Separate refactoring from feature work. A PR that refactors existing code and adds new behavior is two PRs — say so in the review. Small cleanups (variable renaming) can be included at reviewer discretion.
PR Descriptions
Review the PR description itself — it becomes the merge commit and stands alone in version control history.
Title: Short, imperative, standalone. "Delete the FizzBuzz RPC" not "Deleting the FizzBuzz RPC." Must be informative enough that someone searching history can understand the change without reading the diff.
Body: What is changing and why. Include context, decisions, and reasoning not visible in the code itself. Link to bug numbers, benchmark results, or design docs where relevant. Acknowledge approach shortcomings when they exist.
Anti-patterns: "Fix bug," "Fix build," "Add patch," "Moving code from A to B," "Phase 1," "Add convenience functions."
A weak description is a legitimate review finding — raise it in the summary body, not as an inline comment.
Review Process
Step 2: Understand the Context
Before looking at code, understand the intent:
- What is this PR trying to accomplish?
- What issue or spec does it implement?
- What is the expected behavior change?
Step 3: Review the Tests First
Tests reveal intent and coverage:
- Do tests exist for the change?
- Do they test behavior (not implementation details)?
- Are edge cases covered?
- Do tests have descriptive names?
- Would the tests catch a regression if the code changed?
Step 4: Review the Implementation
Walk through the code with the five axes in mind:
For each file changed:
1. Correctness: Does this code do what the test says it should?
2. Readability: Can I understand this without help?
3. Architecture: Does this fit the system?
4. Security: Any vulnerabilities?
5. Performance: Any bottlenecks?
Step 5: Categorize Findings
Prefix every finding with its severity so the author knows what's required vs optional:
Prefix
Meaning
Author Action
(no prefix)
Required change
Must address before merge
Critical:
Blocks merge
Security vulnerability, data loss, broken functionality
Nit:
Minor, optional
Author may ignore — formatting, style preferences
Optional: / Consider:
Suggestion
Worth considering but not required
FYI
Informational only
No action needed — context the author lacks
This prevents authors from treating all feedback as mandatory and wasting time on optional suggestions. The prefix goes at the start of the inline comment body so it's visible in the GitHub UI without expanding.
Every inline comment must be actionable or tell the author something they don't already know. An inline comment is a notification, a thread someone has to read and resolve, and a permanent line in the PR's history. It has to earn that.
Never post praise as an inline comment. No "great context here," no "nice test coverage," no "good use of X." The author knows what they wrote; complimenting it back to them is noise that buries the findings that matter. This is the single most common way an AI review degrades into spam:
❌ FYI: Great context in this comment block. Explaining the 'why' alongside the
test assertions makes this much easier to maintain and prevents accidental
regressions.
❌ Nit: Nice work extracting this into a helper — much more readable.
❌ FYI: This correctly handles the nil case.
None of those change anything. Delete them.
FYI is not a praise slot. It's for context the author genuinely lacks and can't get from the diff — "this endpoint is rate-limited to 10 req/s upstream, so the retry loop will trip it," "we hit this exact race in #482." If you can't name what the author would do differently knowing it, it isn't an FYI.
If the PR is genuinely good, say so once — in the summary body, in one line. "Clean change, tests cover the branches that matter" is worth more than five scattered compliments, and it costs the author nothing to read.
Before submitting, re-read every inline comment and drop any that: only compliments, only restates what the code does, or says something the diff already makes obvious. A review of 3 real findings beats a review of 3 real findings buried in 9 pleasantries.
Step 6: Verify the Verification
Check the author's verification story:
- Do CI checks pass? (gh pr checks)
- What tests were run?
- Was the change tested manually?
- Are there screenshots for UI changes?
- Is there a before/after comparison?
Posting the Review
The review is delivered as one GitHub review containing every inline comment plus the summary body. Never post findings as a series of separate issue comments — that spams the PR and loses the file/line anchoring.
Event type
Always submit with event: "COMMENT". Never APPROVE, never REQUEST_CHANGES. The verdict is stated in the review body text; GitHub's merge gate stays under human control.
Anchoring inline comments
Each inline comment needs:
Field
Value
path
Repo-relative file path, exactly as it appears in the diff
line
Line number in the file at the given side — not a diff offset
side
RIGHT for added/unchanged lines (the new file), LEFT for removed lines (the old file)
start_line + start_side
Optional — for a multi-line comment spanning a range, with line as the end
body
The finding, severity-prefixed
The hard constraint: line must fall inside a hunk present in the PR diff. GitHub rejects the entire review with 422 Unprocessable Entity if any single comment points outside the diff. Derive valid ranges from the hunk headers:
@@ -oldStart,oldCount +newStart,newCount @@
RIGHT-side comments must land in [newStart, newStart + newCount - 1]; LEFT-side in [oldStart, oldStart + oldCount - 1]. Verify every comment against these ranges before submitting.
If a finding can't be anchored — it's about a file not in the diff, about something missing, or about the change as a whole — put it in the summary body under "General findings" with a path:line reference in text. Do not force it onto an unrelated line just to make it inline.
Suggested fixes render as applyable suggestion blocks when the fenced block is tagged suggestion and its content replaces exactly the commented line range:
Nit: this shadows the outer `err`.
```suggestion
if writeErr := w.Flush(); writeErr != nil {
```
Submitting with gh
gh pr review cannot attach inline comments — use the API with a JSON payload. Write the payload to a file rather than inlining it, so multi-line bodies and backticks survive the shell:
cat > /tmp/pr-review-NUMBER.json <<'JSON'
{
"commit_id": "HEAD_SHA",
"event": "COMMENT",
"body": "## Review summary\n\n...",
"comments": [
{
"path": "src/auth/session.go",
"line": 84,
"side": "RIGHT",
"body": "**Critical:** the session token is logged in full here..."
},
{
"path": "src/auth/session.go",
"start_line": 120,
"start_side": "RIGHT",
"line": 128,
"side": "RIGHT",
"body": "**Consider:** this block duplicates `refreshToken` above..."
}
]
}
JSON
gh api repos/OWNER/REPO/pulls/NUMBER/reviews \
--method POST \
--input /tmp/pr-review-NUMBER.json
Generate that JSON with a script (jq, or a small heredoc'd Python) when there are many findings — hand-escaping newlines in long bodies is where this breaks. commit_id is optional but pin it to the head SHA so comments stay attached to the revision you actually reviewed.
Fallbacks when gh is unavailable or unauthenticated
In priority order:
GitHub MCP tools, if present in the session — mcp__github__create_pull_request_review takes the same body / event / comments[] shape in one call. Other servers may expose a pending-review flow instead (create pending review → add comments one at a time → submit); if so, always submit at the end, or the review is left in a pending state only you can see.
Raw REST with a token: curl -X POST -H "Authorization: Bearer $GITHUB_TOKEN" -H "Accept: application/vnd.github+json" https://api.github.com/repos/OWNER/REPO/pulls/NUMBER/reviews -d @/tmp/pr-review-NUMBER.json
No write access at all — say so plainly, output the review in the terminal instead, and hand over the exact command to post it. Don't pretend it landed.
Handling failures
Symptom
Cause
Fix
422 — "line must be part of the diff"
A comment points outside a diff hunk
Re-check that comment against hunk ranges; move it to the summary body if it can't anchor
422 — "pull_request_review_thread.path is invalid"
Path doesn't match the diff exactly (case, leading ./, renamed file)
Copy the path verbatim from gh api .../files
403 / 404 on POST
No write access, or a fine-grained token missing pull_requests: write
Fall back per above
422 on APPROVE for your own PR
GitHub forbids self-approval
Not applicable here — this skill always uses COMMENT
A 422 rejects the whole review, so nothing is posted twice. Fix the offending comment and resubmit the full payload. If the same payload partially succeeded (it doesn't, but if in doubt), check gh api repos/OWNER/REPO/pulls/NUMBER/reviews before resubmitting to avoid a duplicate review.
The summary body
The review body is the overall assessment. Keep it tight — the detail lives inline.
The first line must attribute the review to the model that produced it — model name, and reasoning effort when that's known. This tells the author what reviewed their code and how hard it thought about it. Use the model you are actually running as, not a placeholder:
_Reviewed by Claude Opus 5 (reasoning effort: high)._
If the effort level isn't available to you, state the model alone — don't invent a level. Structure:
_Reviewed by <modelname> (reasoning effort: <level>)._## Review summary
One or two sentences: what this PR does and the overall read on it.
**Verdict:** Approve / Request changes / Needs discussion — with the one-line reason.
### Blocking-`src/auth/session.go:84` — session token written to logs at info level
### Should fix-`src/api/list.go:210` — N+1 query, one lookup per row; ~50ms added per 100 items
### Optional- 3 nits left inline.
### Coverage
Reviewed all 12 changed files. CI: 2 checks passing, 1 failing (`test-integration`).
Tests: added for the happy path; no test for the expired-token branch.
State the verdict in the body even though the GitHub event is always COMMENT — that's how the reader gets your actual call.
Re-reviewing After Changes
When asked to re-review a PR you already reviewed:
Read your prior review and its threads: gh api repos/OWNER/REPO/pulls/NUMBER/comments --paginate
Diff only what's new since your last review: gh pr diff NUMBER --repo OWNER/REPO against the SHA you previously pinned
Note which prior findings are resolved, which are outstanding, and which were addressed differently than suggested
Reply in the existing thread for a still-open finding rather than opening a new one: gh api repos/OWNER/REPO/pulls/NUMBER/comments/COMMENT_ID/replies -f body='...'
Post a new review only for new findings, and open the body with the status of the previous round
Don't re-report a finding the author already declined with a stated reason. Note it as resolved-by-discussion and move on.
Dead Code Hygiene
Check the PR for orphaned code:
Identify code that is now unreachable or unused as a result of this change
List it explicitly in the review body
Ask, don't demand deletion — the author may know of a caller you can't see
DEAD CODE IDENTIFIED:
- formatLegacyDate() in src/utils/date.ts — replaced by formatDate()
- OldTaskCard component in src/components/ — replaced by TaskCard
- LEGACY_API_URL constant in src/config.ts — no remaining references
→ Safe to remove these in this PR?
Review Speed
Slow reviews block entire teams. The cost of context-switching to review is less than the waiting cost imposed on others.
Respond within one business day — this is the maximum, not the target
Prioritize fast individual responses over quick final approval. Quick feedback reduces frustration even if multiple rounds are needed
Large PRs: ask the author to split them rather than reviewing one massive changeset
Handling Disagreements
When resolving review disputes, apply this hierarchy:
Technical facts and data override opinions and preferences
Style guides are the absolute authority on style matters
Software design must be evaluated on engineering principles, not personal preference
Codebase consistency is acceptable if it doesn't degrade overall health
Don't accept "I'll clean it up later." Experience shows deferred cleanup rarely happens. Require cleanup before merge unless it's a genuine emergency. If surrounding issues can't be addressed in this PR, ask for a filed issue with an assignee.
Honesty in Review
The review is public and attributed. That raises the bar, it doesn't lower it:
Don't rubber-stamp. "LGTM" without evidence of review helps no one.
Don't soften real issues. "This might be a minor concern" when it's a bug that will hit production is dishonest.
Quantify problems when possible. "This N+1 query will add ~50ms per item in the list" is better than "this could be slow."
Push back on approaches with clear problems. Sycophancy is a failure mode in reviews. If the implementation has issues, say so directly and propose alternatives.
Comment on code, not people. Reframe personal critiques to focus on the code itself — this matters more in a public thread than in a terminal.
Accept override gracefully. If the author has full context and disagrees, defer to their judgment.
Don't pad. Five inline nits on a clean PR is noise. If the PR is good, say it's good and post the few things that matter.
Dependency Discipline
Part of PR review is dependency review. If the PR touches a manifest or lockfile:
Does the existing stack solve this? (Often it does.)
How large is the dependency? (Check bundle impact.)
Is it actively maintained? (Check last commit, open issues.)
Does it have known vulnerabilities? (npm audit, govulncheck, or the language equivalent)
What's the license? (Must be compatible with the project.)
Rule: Prefer standard library and existing utilities over new dependencies. Every dependency is a liability.
The Review Checklist
## Review: OWNER/REPO#NUMBER — [PR title]### Context- [ ] PR was explicitly provided (never inferred from the branch)
- [ ] I understand what this change does and why
- [ ] I read beyond the diff where context was needed
### Correctness- [ ] Change matches spec/issue requirements
- [ ] Edge cases handled
- [ ] Error paths handled
- [ ] Tests cover the change adequately
### Readability- [ ] Names are clear and consistent
- [ ] Logic is straightforward
- [ ] No unnecessary complexity
### Architecture- [ ] Follows existing patterns
- [ ] No unnecessary coupling or dependencies
- [ ] Appropriate abstraction level
### Security- [ ] No secrets in code
- [ ] Input validated at boundaries
- [ ] No injection vulnerabilities
- [ ] Auth checks in place
- [ ] External data sources treated as untrusted
### Performance- [ ] No N+1 patterns
- [ ] No unbounded operations
- [ ] Pagination on list endpoints
### Verification- [ ] CI checks reviewed
- [ ] Verification story documented in the PR
### Delivery- [ ] Review body opens with the model attribution line (model + effort)
- [ ] Every inline comment anchored to a line inside a diff hunk
- [ ] Every finding severity-prefixed
- [ ] No inline comment is praise, and none merely restates the code
- [ ] Every inline comment is actionable or tells the author something new
- [ ] Unanchorable findings moved to the summary body
- [ ] Submitted as ONE review with `event: "COMMENT"`- [ ] Review URL reported back to the user
Common Rationalizations
Rationalization
Reality
"It works, that's good enough"
Working code that's unreadable, insecure, or architecturally wrong creates debt that compounds.
"I'll just infer which PR they meant"
Reviewing the wrong PR posts public comments on someone else's work. Ask.
"The diff is enough context"
A hunk shows what changed, not whether it's correct. Read the file and its callers.
"I'll post each finding as its own comment"
That spams the PR with notifications. One review, many inline comments.
"Positive feedback is encouraging"
An inline compliment is a notification and a thread to resolve. One line in the summary does the job without burying the real findings.
"It's only an FYI, it's harmless"
Every inline comment costs the author attention. If it doesn't change what they'd do, it's noise.
"AI-generated code is probably fine"
AI code needs more scrutiny, not less. It's confident and plausible, even when wrong.
"The tests pass, so it's good"
Tests are necessary but not sufficient. They don't catch architecture problems, security issues, or readability concerns.
"422 means it partly worked"
A 422 rejects the entire review. Nothing posted. Fix and resubmit the whole payload.
Red Flags
Reviewing a PR the user didn't name
Posting the review to the terminal instead of to the PR
Inline comments that praise the code instead of flagging something
FYI used as a compliment slot rather than for context the author lacks
Comments that restate what the code plainly does
Findings with no path/line that could easily have had one
A review body with no model attribution, or one naming a model you aren't
Review that only checks if CI passed (ignoring the other axes)
"LGTM" without evidence of actual review
Review comments without severity labels — makes it unclear what's required vs optional
Security-sensitive changes without security-focused review
Large PRs that are "too big to review properly" (ask for a split)
No regression tests with a bug-fix PR
Claiming full coverage of a PR you only skimmed
Verification
After posting:
The review exists on the PR — confirm with gh api repos/OWNER/REPO/pulls/NUMBER/reviews --jq '.[-1] | {id, state, user: .user.login}'
Inline comments landed — gh api repos/OWNER/REPO/pulls/NUMBER/comments --jq 'length'
No duplicate review was created by a retry
The review URL was reported to the user, with a one-paragraph summary of what was posted
See Also
For detailed security review guidance, see references/security-checklist.md
For performance review checks, see references/performance-checklist.md
code-review-and-quality — the same five-axis standard for local, uncommitted changes
address-pr-review-comments — the other side: fetching, fixing, and resolving review threads on a PR
gh — conventions for GitHub CLI usage in this setup
